Semi-Supervised Deep Expectation-Maximization For Low-Dose Pet-Ct

Reducing the dose of ionizing radiation underlying combined imaging with positron emission tomography (PET) and computed tomography (CT) typically leads to reduced image quality. We propose a novel variational deep-neural-network (DNN) framework for image quality enhancement of low-dose PET-CT images, relying on Monte-Carlo expectation maximization. Unlike existing DNN-based training that pairs low-dose PET-CT images with their corresponding high-dose versions, we propose a semi-supervised learning framework that enables learning using a small number of high-dose images. We propose a robust and uncertainty-aware loss motivated by a heavy-tailed generalized-Gaussian distribution on the residuals between the DNN output and the PET-CT data, aiding our semi-supervised learning scheme. Results on a publicly available dataset show the benefits of our framework, quantitatively and qualitatively, over existing methods.